SCIT: Testing Causal Cache Carriers in Latent Chain-of-Thought Models

TL;DR - SCIT is a causal testing protocol for identifying which transformer cache components carry computations in latent chain-of-thought models. It finds that arithmetic reasoning in tested GPT-2 checkpoints primarily transfers through value-cache suffix trajectories, while larger or non-arithmetic models can use different cache regions.

• SCIT patches exact cache segments between source and recipient examples, combining sufficiency, necessity, component-split, corruption, and control tests.
• CODI-GPT2 showed sufficiency-and-necessity evidence that late value-cache suffixes carry counterfactual arithmetic computations.
• A Sim-CoT-style GPT-2 reproduced the sufficiency pattern, but lacked enough matched-corruption evidence to establish necessity.
• Carrier mechanisms varied with model competence and task: some cells used latent-tail values/KV, while competent 8B and repaired non-arithmetic cells relied on prompt-prefix or full-cache K/V.